This M.Tech dissertation bridges the gap between deep learning and green building domains, contributing to the larger goal of sustainable urban development. The study focuses on using advanced semantic segmentation techniques to identify key façade components in existing buildings, enabling retrofitting with green features like vegetation, vertical gardens, or energy-efficient design.
The construction sector significantly contributes to environmental degradation. Green buildings offer a sustainable solution, but traditional tools often lack the precision to retrofit existing buildings with green standards. In this work, a deep learning model has been developed to segment building façades, helping to analyze and visualize suitable zones for green implementations like planting systems, energy panels, and ventilation features.
- Modern urban areas demand transformation into eco-friendly spaces.
- Building exteriors provide untapped potential for hosting vertical greenery.
- Existing green building tools are tailored for new constructions, but this research focuses on already-built structures, providing scalable solutions to upgrade them with sustainable features.
- With accurate façade segmentation, we can target features like balconies, windows, or pillar that could support vegetation, improve air quality, and boost aesthetic value.
The following objectives were formulated based on a thorough literature review and gap identification:
- Develop a Deep Learning-based robust U-Net Model for accurate building facade segmentation.
- Enhancement of data quality by using pre-processing techniques, edge detection and creation of a new dataset.
- Analyse segmented components to enhance vegetation and sustainability to choose 5 most relevant components for labelling section.
- Base Dataset: CMP Facade Dataset
- Enhanced Dataset: CED Dataset
- Binary Mask Splitting technique was applied to achieve fine-grained separation of labels.
- Canny Edge Detection was applied to improve edge clarity.
- Custom data augmentation was employed to address class imbalance and increase the size of the dataset.
UNet Model
The U-Net architecture is a convolutional neural network designed for image segmentation tasks. It has an encoder-decoder structure, where the encoder captures context by down-sampling the input image, and the decoder reconstructs the image at a higher resolution. Skip connections between corresponding layers in the encoder and decoder help preserve spatial information, which is crucial for accurate segmentation. U-Net’s ability to deliver pixel-level accuracy makes it ideal for building segmentation tasks, ensuring that even small but important features are correctly identified, which is vital for green building assessments.
A custom UNet model was built and enhanced with a Spatial Multiplicative Cross-Attention Mechanism to focus on semantically relevant regions. The network follows a typical encoder-decoder structure with:
- Convolution → DropOut → BatchNorm → Activation → Pooling
- Attention mechanism applied in the upsampled feature maps and skip tensor values
- Optimizer:
Adam, Loss:Categorical Focal Loss - Metrics:
Accuracy,Mean IoU,Precision,Recall,F1-Score
UNet architecture with Spatial Attention components.
- Training was performed for 9 fits (epochs) using the enhanced CED dataset.
- The final model demonstrated consistent improvement in segmentation accuracy and class precision: 98.1% accuracy, 87% mean Intersection over Union score and 93% f1-Score.
- Visual comparisons of predictions indicate accurate boundary detection and effective component differentiation.
- The model successfully segments building images to highlight façade components suitable for green retrofit.
The final labelled section consists of five façade classes most aligned with green building standards:
| Label | BGR Color Code | Applicability |
|---|---|---|
| Window | (255, 85, 0) |
Placement for air-purifying plants |
| Balcony | (85, 255, 170) |
Candidate zones for vertical gardens, hanging plants and small vegetation |
| Pillar | (0, 0, 255) |
For facade-supported green walls or climbing plants |
| Molding | (0, 85, 255) |
Vertical Planters, hanging plants or ornamental plants |
| Sill | (0, 170, 255) |
Living Wall System where plants substrate |
This research highlights how deep learning, when thoughtfully integrated with domain-specific knowledge, in this case, green building principles; can address real-world challenges such as environmental sustainability and urban infrastructure planning. By blending traditional architectural understanding with modern computer vision techniques, we not only enhance the precision of building analysis but also contribute to more informed, eco-conscious decision-making. Moving forward, such interdisciplinary approaches will be vital in designing resilient, intelligent systems that serve both technological progress and planetary well-being.